Abstract: Objectives: Under cloud computing environments, remote sensing cloud computing platforms typically achieve task orchestration by flexibly invoking and freely combining spatiotemporal operators through code scripts, thereby providing diversified remote sensing analysis capabilities. The complexity and uncertainty of geospatial problems prompt researchers to decompose problems into several easily solvable subproblems for step-by-step solving. However, existing task orchestration methods primarily focus on holistic problem modeling, making it difficult to support phased local construction and iterative adjustment of spatiotemporal tasks. Furthermore, in interactive scenarios, they tend to generate numerous task flows with similar logic that are submitted to the cloud for execution, resulting in redundant computations. Methods: A spatiotemporal task interactive computing framework has been designed. To address the demand for efficient computing resource utilization, an incremental computing method for spatiotemporal task targeted reconstruction is proposed. At the framework level, spanning the entire lifecycle of spatiotemporal task construction, evolution, and execution, the orchestration process is top-down divided into three orchestration spaces: the abstract overview layer, the logical instance layer, and the physical execution layer. A formal expression model and dynamic construction strategy for spatiotemporal tasks, adaptable to user interaction behaviors, are designed. At the methodological level, we precisely identify newly added/deleted/modified vertices and their associated dependency edges through graph structural difference analysis. By tracing propagation paths using breadth-first search, we locate the boundaries of subgraphs affected by vertex changes, extracting logically changed subgraphs and logically unchanged subgraphs to reconstruct task logic. For logically unchanged subgraphs, we directly match them with historical task flow graphs and reuse their execution results. For logically changed subgraphs, we add input data vertices and associate them with execution results from unchanged subgraphs. This constructs a new spatio-temporal task flow graph, forming a local spatio-temporal task execution plan. Results: The implementation was based on the Open Geospatial Engine (OGE) platform. Through specific case studies, the feasibility of the method was validated. Quantitative results indicate that compared to traditional fullprocess submission methods, the proposed approach reduces computational redundancy by 40%-60% and decreases average task execution time by approximately 35%. Conclusions: The incremental computation method based on task graph structure-targeted reconstruction demonstrates broad applicability across tasks with varying computational characteristics. It effectively reduces computational redundancy and enhances task execution efficiency for diverse interactive behaviors, such as vertex insertion/deletion and dependency relationship adjustments. Furthermore, when handling large-scale data processing tasks, this approach holds significant value in reducing memory resource consumption and improving system stability.